How large language models (llms) like chatgpt will impact medicine

3 March 2023
Within just two months after being released, ChatGPT gained 100 million users, while it took TikTok, the former leader, nine months. This results from the incredible capabilities of large language models (also known as conversational AI) and chat's practical application. ChatGPT enables users to get answers within seconds. Most importantly, the text is natural-sounding with a consistent form and language style. AI models are refined enough to create new content based on available knowledge, including poems, recipes, music, and even paintings. Similar models are being developed for healthcare. They enable patients to accurately asses their health and obtain health recommendations based on state-of-the-art scientific knowledge. Physicians and nurses will soon be able to "ask" an electronic health record instead of conducting a manual search.

Med-PaLM, ChatGTP-like chatbot for medical purposes

At the end of 2022, Google and DeepMind launched Med-PaLM, an AI-powered chatbot. The purpose of the system is to generate accurate and safe responses regarding medical and health issues. This new tool – similar to ChatGPT – is based on Large Language Models (LLMs). LLMs are models processing natural language by artificial intelligence. The aim is to understand inquiries and produce plain-language text responses. Regarding Med-PaLM, the datasets used to generate answers concern only medical issues and come from scientific research. Med-PaLM was trained using seven sets of questions and answers covering professional medical exams, scientific research, and healthcare-related web searches. A study published by Google and DeepMind, "Large Language Models Encode Clinical Knowledge," shows that the model, once improved, can be used for clinical purposes. Researchers admit that although its performance is "encouraging," Med-PaLM remains inferior to physicians' accuracy. Out of 140 questions assessed in the study, 92.9% of clinicians' answers were consistent with scientific knowledge. For the AI model, it was only 61.9%. Incorrect retrieval of information was seen in 16.9% of answers generated by Med-PaLM (for human clinicians, it was less than 4%). A similar disproportion was observed for incorrect reasoning (10% vs. 2%) as well as inadequate or inaccurate response (18.7% vs. 1.4%). However, the AI model is a fast learner. ChatGPT-3, the current version of the model, draws knowledge from a total of 175 billion parameters gathered from approx. 10 million web pages. The following versions are expected to be hundreds of times more powerful. This would enable it to generate videos, for example. A similar pace can be expected in medicine.

Five applications of large language models in medicine

  •  Evidence-based medicine A physician will be able to quickly find new scientific research useful for a patient's treatment. For instance, it will take only to ask the right question: "search for the latest research on the treatment for vancomycin-resistant enterococci infection in a 70-year-old patient with diabetes". Making the answer more precise will be possible by asking about specific recommendations. Doctors will be able to ask directly about current medical advice, the effectiveness of therapy, and potential complications and get a response within seconds.
  • Looking for the answer in EHR. A physician will get information that manually would take hours to find. For instance: "check the results for leukocytes levels over the last five years and show them on a timeline," or "check if there are any new studies from the point of view of the patient's medical record and the current treatment" or "which drug for patients with parameters like patient X is most effective."
  • Seeking a second opinion. The AI health models will serve as a primary specialist consultations for general practitioners. In the same way, as AI already interprets medical imaging, new models will aid the diagnostics of complex clinical cases and rare diseases. Again, the place of treatment will not matter – a patient will receive the same quality of care regardless of whether it is a modern medical center of a big university hospital or a small doctor's office in a village.
  • Preliminary diagnosis. The fact that Google develops the Med-PaLM model is not random. Google is currently the world's biggest medical encyclopedia. Its search engine receives about 70,000 health-related inquiries per minute. Unfortunately, the answers are often misleading as they are based on popularity rather than the quality or credibility of the content. Google knows that the competition, in the form of symptom checkers and AI-driven chats, is growing.
  • Personal health advisors. "What dietary supplements should people suffering from X use at the age of Y when taking Z medications." Medical chatbots in the form of applications and robots will observe our daily lives, analyze data, compare them to the knowledge available on the Internet, correct unhealthy habits, give advice on better ways to maintain good mental and physical condition, eat properly, and do sports. What's more, you won't need to know what to ask – chatbots will be the ones to suggest and guide actively.
In a few years, medical AI models will become trusted assistants of physicians, relieving them of their keyboards and searching for data in a pile of information.